What happened
Draw Things announced on October 28, 2026 that it was releasing Local Code to the public and making the entire application open source. The company had discussed the application during a public beta a few weeks earlier. Local Code is available for download from getlocalcode.app, while a macOS App Store release is described as coming soon. The project repository is available at GitHub under the Draw Things organization.
The product is positioned as a general-purpose local agent for a Mac rather than only a coding assistant. Draw Things says it can write or review Python and JavaScript, generate images, videos, infographics, slides, and explainer videos, and work with documents and spreadsheets. Its bundled environment includes Python and draw-things-cli for media work. The post also says the agent can visit a website, log in with a user's credentials, and make an appointment. It can connect to Photos, Messages, Mail, Notes, and Calendar to summarize content or edit and reorganize it where the relevant application supports that action.
Local use is designed not to require external installation steps, a sign-in, API keys, or external MCP servers. Models are selected and downloaded through the application. That packaging is significant: the model, agent interface, execution environment, and application integrations are presented as one local product boundary. The post frames the privacy benefit in direct terms: prompts and files need not be sent to a cloud AI provider when the chosen model runs locally.
Draw Things lists three quantized model packages with very different hardware envelopes. Qwen 3.8 27B is a 10 GiB package intended for Macs with at least 16 GiB of RAM. DeepSeek 4 Flash 0731 is a text-only 304B-parameter model in an 89 GiB package for Macs with 48 GiB or more. The largest option, DeepSeek 4.1 Flash 1017, is described as a 763B-parameter model quantized into a 376 GiB package. Draw Things says it can run with about 36 GiB of resident memory, making it suitable for Macs with 64 GiB or more, and reports about 350 tokens per second for prompt processing and about 18 tokens per second for generation. The company says its own benchmarks closely matched officially reported results and that its two-bit S quantization retained intelligence without regression. It also says the format uses M5 Ultra Neural Accelerators to improve prompt processing and time to first token.
What the archive adds
This is a new event, but it intersects with the earlier note that OpenCode made DeepSeek V4.1 Flash available with a 200K context cap. That story concerned hosted access through a coding-agent platform; Local Code instead packages a local runtime and asks the user to provide a suitably capable Mac.
Trend signal
Local Code supports the pattern that harness architecture outlives model choice. Its claim is not only that a large model can run locally, but that models of different sizes can sit behind the same agent surface with browser, shell, code, document, and media capabilities. The practical product boundary is the runtime around the model.
Why now
The timing follows a public beta that began only weeks before the release. Draw Things is moving from an early public test to a downloadable, open-source application while promising a later Mac App Store distribution path. The announcement also arrives with several model sizes, allowing the same application to target both 16 GiB Macs and much larger machines rather than depend on one hardware class.
Connections to other events
The earlier OpenCode note illustrates the other side of access to DeepSeek V4.1 Flash: a hosted coding environment with a stated 200K context cap. Local Code is not presented as a hosted offer. Its trade-off is local execution and data locality in exchange for substantial storage and memory requirements for the largest packages.
Why it matters
For teams building with AI, Local Code is useful less as proof that every employee should run a 763B-parameter model and more as a reference architecture for local agent work. It combines model selection, tool execution, local application access, and source availability in a single distribution. That makes it possible to inspect how a desktop agent exposes capabilities such as file-adjacent applications, browser actions, code execution, and media generation.
The release also makes the infrastructure trade-off explicit. Privacy and reduced dependency on a cloud provider do not remove operating constraints; they move them into hardware capacity, downloaded package size, performance, and the permissions granted to the desktop application. Teams handling sensitive material can evaluate whether a local runtime is appropriate for particular workflows, while keeping the largest-model claims separate from the needs of a routine code-review or document task.
What happens next
- More local-agent evaluation will focus on tool permissions and hardware envelopes, not only benchmark claims. (3 months, 65%). Watch for: published guidance or controls for application access, browser actions, and local data.
- A Mac App Store release will broaden installation access but may clarify which capabilities fit store distribution. (6 months, 60%). Watch for: the announced App Store listing and any stated capability differences.
Limits and uncertainty
The source is a product announcement from Draw Things. Its performance figures, benchmark comparisons, suitability thresholds, and claims about intelligence retention are the company's own statements; it does not provide an independent evaluation in the supplied material. It also does not establish how reliably Local Code completes complex browser, document, or application tasks in real production workflows. “Local” does not itself describe the application's permission model, isolation boundaries, network behavior during every task, or the security implications of asking an agent to log in to websites. Finally, the announcement gives no comparison of cost, latency, or quality against cloud coding agents for a defined workload.
Skill to build now
Map one workflow that handles sensitive source files or internal documents and separate the model decision from the runtime decision. List the exact capabilities the agent would need: file access, shell execution, browser login, document editing, or application integrations. Then define which of those permissions must be available locally and which can remain in a cloud service. This turns “run locally” from a privacy slogan into an evaluable design choice.
Reader takeaway
Local Code shows what an on-device agent product can look like when the model is only one component: the useful unit is a packaged runtime with tools, integrations, and a local execution path. Its largest configurations are demanding, but the open-source release gives builders a tangible system to inspect when designing private agent workflows.
Open question
Can a local desktop agent provide enough observable permission controls and task verification to make access to personal applications safer than sending the same work to a cloud agent?